SERIES
PAPER
DISCUSSION
IZA DP No. 8287
The Effect of Conflict History on Cooperation
Within and Between Groups: Evidence from a
Laboratory Experiment
Gonne Beekman
Stephen L. Cheung
Ian Levely
June 2014
Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
The Effect of Conflict History on
Cooperation Within and Between Groups:
Evidence from a Laboratory Experiment
Gonne Beekman
Wageningen University
Stephen L. Cheung
University of Sydney
and IZA
Ian Levely
Charles University in Prague
Discussion Paper No. 8287
June 2014
IZA
P.O. Box 7240
53072 Bonn
Germany
Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: [email protected]
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The IZA research network is committed to the IZA Guiding Principles of Research Integrity.
The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 8287
June 2014
ABSTRACT
The Effect of Conflict History on Cooperation Within and
Between Groups: Evidence from a Laboratory Experiment *
We study cooperation within and between groups in the laboratory, comparing treatments in
which two groups have previously been (i) in conflict with one another, (ii) in conflict with a
different group, or (iii) not previously exposed to conflict. We model conflict using an intergroup Tullock contest, and measure its effects upon cooperation using a multi-level public
good game. We demonstrate that conflict increases cooperation within groups, while
decreasing cooperation between groups. Moreover, we find that cooperation between groups
increases in response to an increase in the efficiency gains from cooperation only when the
two groups have not previously interacted.
JEL Classification:
Keywords:
C92, D64, D74, H41
within- and between-group cooperation, inter-group conflict, group identity,
multi-level public good experiment, Tullock contest, other-regarding preferences
Corresponding author:
Stephen L. Cheung
School of Economics
Merewether Building H04
The University of Sydney
NSW 2006
Australia
E-mail: [email protected]
*
We thank Klaus Abbink, Daniel Friedman, Simon Gächter, and Werner Güth; seminar audiences at
the University of New South Wales, Monash University and University of Sydney; and participants in
the Trento Summer School on Evolution of Social Preferences in July 2011, the Wageningen
International Conference on Cooperation or Conflict in May 2013, the Workshop on Competition
between Conflict and Cooperation in Freiburg in June 2013, the International Meetings of the
Economic Science Association in Zurich in July 2013, the Australia and New Zealand Workshop on
Experimental Economics in Brisbane in August 2013, the Asia-Pacific Meetings of the ESA in
Auckland in February 2014, and the Thurgau Experimental Economics Meeting on Cooperation and
Competition Within and Between Groups in May 2014.
1. Introduction
In many interesting settings, a period of conict or competition between groups is followed by the opportunity for mutually benecial cooperation between the same groups.
Examples include the formation of a coalition government following an election, the integration of work teams following a corporate merger, and the reunication of a nation
after a period of civil conict.
In such situations, individuals are faced with a choice
between acting in their own self-interest, in the parochial interests of their in-group, or
in the collective interests of all parties.
If, as a result of the prior history of conict, individuals are reluctant to cooperate with
members of an out-group, the result may be substantial eciency losses to society
as a whole.
Yet, at the same time, a shared experience of conict may also reinforce
cooperative norms among members of an in-group.
To give an extreme example: two
decades after the wars in the Balkans, Muslims and Christians in Bosnia have established
separate schools and even separate re departments (Brunwasser, 2011).
This clearly
illustrates these groups' preference to invest in local public goods that only benet
members of their in-group, as opposed to global public goods that benet all parties.
There are several possible reasons why conict may inhibit subsequent cooperation between groups. Firstly, the underlying reasons for the conict could also have an eect on
cooperation. Secondly, conict could create or deepen in-group identity, strengthening
other-regarding preferences toward in-group members and making it more attractive to
cooperate within groups. Finally, conict may create animosity towards out-group members, eroding other-regarding preferences towards out-group members and making it less
attractive to cooperate between groups.
In this paper, we report a laboratory experiment designed to study how cooperative behavior, both within and between groups, is inuenced by the group members' experience
of a prior phase of conict. In particular, we compare levels of within- and between-group
cooperation in the situation described above where the
same
two groups were previ-
ously in conict to a comparable situation in which each group previously experienced
conict involving a
dierent
out-group, as well as when group members have
no
prior
experience of conict. We vary this group matching on a between-subjects basis.
Since exposure to conict is exogenous and randomly-assigned in our experiment, we can
set aside the rst explanation namely that conict and cooperation share some common
and deeper source. Our group matching manipulations then enable us to disentangle the
latter two mechanisms, to independently identify the eects of a shared experience of
conict upon other-regarding preferences toward members of the in- and out-groups.
2
Our instrument for measuring cooperation within- and between-groups is a multi-level
public good (MLPG) game (Blackwell and McKee, 2003). In this game, all individuals
have an endowment which they can retain for private consumption, contribute to a local
public good that benets only members of the in-group, and/or a global public good that
benets members of both the in- and out-groups. Our conict manipulation takes the
form of an inter-group version of the Tullock rent-seeking contest (Tullock, 1980; Abbink
et al., 2010). The Tullock game thus models a prior phase of inter-group conict which
is followed by a subsequent opportunity for cooperation in the form of the MLPG game.
Previous studies of the MLPG game typically nd that contributions to the global public
good are increasing in the relative return on the global account compared to the local
one (Blackwell and McKee, 2003; Fellner and Lünser, 2008; Chakravarty and Fonseca,
2013).
However, since the Tullock contest potentially induces a much stronger form
of in- and out-group identity than has previously been considered in this literature, this
responsiveness to eciency considerations may not be robust to a shared history of conict
involving the same out-group. For this reason, we vary the return on contributions to the
global public good as a second dimension of our experiment design.
Our approach thus introduces a number of methodological innovations.
Firstly, we go
beyond standard arbitrary or minimal methods of group formation, by using inter-group
competition in a Tullock contest to instil a much stronger form of induced group identity
forged in conict against another group in the laboratory.
Secondly, through our
manipulations of group matching across the two phases of our experiment, we are able to
disentangle the in- and out-group eects of this conict manipulation upon subsequent
interactions in the MLPG game.
Our results are directionally consistent with a simple model of other-regarding preferences
in which conict increases the utility weight on payos for members of the in-group, while
decreasing the weight on payos of the out-group. We nd that within-group cooperation
increases when groups have a shared history of conict compared to when they play the
MLPG without any prior history, while between-group cooperation diminishes when two
groups have previously been in conict. We nd no signicant response to an increase in
the return to between-group cooperation when there has been a prior history of conict
between the groups which is contrary to the results of previous studies that induce
weaker forms of group identity. On the other hand, when two groups have not previously
interacted (but each has nonetheless previously experienced conict involving a dierent
out-group) we nd a signicant increase in between-group cooperation in response to an
increase in its return which is in line with the results of the previous literature.
The paper proceeds by reviewing the relevant literature in Section 2, before Section 3
sets out our experimental design. Section 4 develops a simple model of other-regarding
3
preferences in the MLPG game, from which we derive hypotheses regarding the eects of
our treatments. Section 5 presents and discusses our results, and Section 6 concludes.
2. Related literature
Our paper contributes to two strands of literature.
First, we make a methodological
contribution to the literature on inducing group identity in laboratory experiments. The
most widely used method of doing so, the minimal group paradigm introduced by
Tajfel (1971) in the social psychology literature, involves forming groups on the basis of
seemingly irrelevant personal characteristics such as preference for a particular abstract
painting by either Klee or Kandinsky and has been found to be sucient to induce
a bias in favor of members of the in-group in many psychological experiments (Tajfel
and Turner, 1979). This method has also been widely applied in economic experiments,
1
although usually in a modied form.
In these studies, assigning group membership
randomly according to number or color, or according to trivial preferences, has not always
been sucient to induce an in-group bias. However, using these procedures in a modied
form, and/or in combination with other methods designed to increase the salience of
group membership, has been found to be eective.
One notable example is Chen and Li (2009), who use the Klee/Kandinsky procedure and
nd that subjects are more likely to choose social-welfare maximizing actions in allocation
games when playing with in-group members. In their setting, shared identity deriving
from a trivial preference for one painter over another is only eective in producing differences in behavior toward in- and out-group members when combined with anonymous
communication with the in-group members during a problem-solving task.
Similarly,
Charness et al. (2007) nd that in-group preferences are stronger when an individual's
choices are observed (passively) by in-group members, whereas arbitrarily labeling groups
and identifying them with colors or numbers is not enough to create an in-group bias in
Prisoner's Dilemma and battle-of-the-sexes games.
Eckel and Grossman (2005) compare the eects of several methods for creating group
identity in a laboratory experiment, comparing cooperation in a public good game played
under various degrees of induced group identity, including arbitrary group identity (in
which groups are formed randomly then labeled by color only), as well as treatments in
which identity is strengthened through joint participation in problem-solving tasks, and
competition in a tournament (in which the group with the highest contribution to the
1 As
Chen and Li (2009) point out, the classic denition of the minimal group paradigm requires that
any decisions made by a subject should not directly aect her own payo. However this condition is
violated in most economic experiments that use similar methods to induce group identity in the lab.
4
public good receives a bonus which is deducted from the losing team's payos). They nd
that in-group cooperation is not aected by the arbitrary or problem-solving treatments,
but is signicantly higher when teams have participated in the tournament.
Our approach is similar to Eckel and Grossman (2005) in that we use competition as
a means of making group identity more salient. However, our design diers in two key
respects. First, we are able to examine and compare the eect of this group identity not
only on in-group cooperation after competition, but also on cooperation with an outgroup. Second, we strengthen group identity through competition that produces a social
loss, which may cause a dierent change in preferences than productive competition.
Our design uses a Tullock contest (Tullock, 1980) in which two parties choose how much
to invest in a (wasteful) competition over a xed resource to increase the salience of
group membership. This is the rst study we are aware of to employ the Tullock game
as a means of inducing group identity in order to study subsequent interactions.
In the Tullock game, although any expenditure on competition is inherently socially
inecient, the equilibrium level of investment is positive. In previous experiments, contest investments have far exceeded equilibrium predictions (see Sheremeta (2013) for a
survey). Moreover, playing the Tullock game in a group environment seems to matter:
Abbink et al. (2010) nd that investments in a Tullock contest are even further above
the equilibrium prediction when the game is played in teams rather than individually.
One possible explanation for this nding is that group membership increases utility gains
from winning the contest, which is consistent with the interpretation that group identity
and competition are intertwined.
Abbink (2012) advocates greater experimental research on the Tullock contest as a natural workhorse for the study of conict. In one such study, Ke et al. (2013) nd that when
a pair competes together as a team against a third individual, any in-group solidarity
formed during this rst stage does not diminish the intensity of subsequent conict when
the team members compete over the spoils in a second Tullock contest.
Halevy et al. (2012) also examine how changes in incentives for intra-group competition
aect subsequent interactions. Individuals can choose to cooperate with their in-group
while simultaneously harming the out-group, and after a set number of rounds some
individuals are given the option to cooperate with their in-group without harming the
out-group. The authors nd that individuals prefer to cooperate with in-group members
without imposing negative externalities on the out-group, and this is true even after a
phase of conict in which in-group cooperation is necessarily associated with negative
externalities, resulting in high rates of harm to the out-group.
5
Our study is similar to these experiments in that we are concerned with how conict between groups aects subsequent interactions. However, we study an environment in which
incentives favor conict in the rst phase, and then observe how this aects cooperation
between groups (as opposed to further competition) in a second phase.
This brings us to the second body of literature to which our paper contributes, on cooperation within and between groups including both groups formed endogenously in
the naturally-occurring world as well as ones formed in the lab. Findings from a number of economic experiments show that group membership leads to more within-group
cooperation (Puurtinen and Mappes, 2009), but decreases between-group cooperation.
Therefore, stronger group identity may lead to eciency gains due to increased withingroup cooperation, but eciency losses associated with reduced out-group cooperation.
Several papers use MLPG experiments to study inter-group interactions, generally nding that although subjects contribute non-zero amounts to the local public good, contributions to the global public good are higher, and responsive to its relative eciency.
Blackwell and McKee (2003) use arbitrary group identity groups are formed randomly
and identied by color and nd weak evidence for in-group preferences, but only when
the average per-capita return to the global public good is no lower than that of the local
public good. Fellner and Lünser (2008) study a similar experiment, identifying groups
by letters and adding a monitoring mechanism to increase the salience of group identity.
They nd that contributions to the global public good are high when it is socially more
ecient than the local public good, but that as cooperation decays over time, subjects
switch toward the local public good.
Chakravarty and Fonseca (2013) assign subjects into groups using the Klee/Kandinsky
procedure, and strengthen group identity using intra-group communication (through a
chat-box on the computer screen, as in Chen and Li (2009)) before an MLPG game in
which the eciency of the local public good is varied across treatments. They nd that
even when the nancial return to investing in the global public good is higher, subjects
invest a considerable part of their endowment in the local (club) good, hence reducing
social eciency. One possible reason for the somewhat stronger results of this study may
be that it induces a more salient form of group identity than the two previous studies.
In addition to laboratory experiments, there are also a number of artefactual eld experiments which measure the eects of naturally-occurring group identity on within- and
between-group cooperation.
In a cross-cultural experiment, Buchan et al. (2009) use an adaption of the MLPG game
in which contributions to the local and global public goods do not directly aect current players, but instead accrue to individuals playing in a subsequent session and
6
nd that increased exposure to globalization at both the individual and national level
increases contributions to a global account that also accrues to citizens of other countries.
The MLPG design allows them to separate preferences for cooperating with foreigners
specically from general variation in preferences for cooperation between countries.
The eld experiment most closely related to our paper is Gumen (2012), who studies a
2
variation of the MLPG game using students from fraternities at a US university.
Groups
of subjects from the same fraternity are matched with an out-group either from the same
fraternity or from a rival fraternity, according to treatment. She nds that when subjects
play with an out-group from the same fraternity, they over-invest in the global public
good.
However, when playing with an out-group from a rival fraternity, they invest
comparatively more in the local public good.
In addition to material incentives, inter-group interactions may be motivated by preferences for cooperation within groups and competition between groups (Hirshleifer, 1995),
which may have developed through evolutionary conditions that involved frequent conict
between small groups (Bowles and Gintis, 2011). Choi and Bowles (2007) hypothesize
that war is instrumental in maintaining and strengthening parochial altruism (increased
altruism towards in-group members coupled with hostility towards out-group members).
In line with this theoretical prediction, recent empirical evidence on the eects of war on
social preferences shows that war leads to more altruism towards neighbors (Voors et al.,
2012), stronger egalitarian norms towards in-group members by children (Bauer et al.,
2014), and more within-group cooperation (Gneezy and Fessler, 2012).
Our study is related to this literature, but focuses on the role that simple conict in the
form competition over a xed resource as distinct from exposure to violence or other
trauma plays in shaping social preferences. We contribute to the discussion of group
identity and cooperation by disentangling how conict aects other-regarding preferences
toward the in- and out-groups, using competition in the Tullock contest as novel means
of inducing group identity in the laboratory.
3. Design
Our experiment consists of two stages: a group Tullock contest and an MLPG game,
both of which are played between two groups of three subjects in each of our treatments.
Groups are formed randomly and anonymously at the beginning of the session, and the
membership of a subject's in-group remains the same throughout the experiment. In
each stage, each group is paired with a second group (the out-group) for ten rounds of
2 In
Gumen's design, the payo function is non-linear, with an interior optimum for a selsh agent.
7
repeated play, with one round of each game randomly selected to count for payment at
the end of the experiment. The identity of the out-group remains constant across all ten
rounds of a given game. However, it may change between games according to treatment.
We use the Tullock contest primarily to manipulate subjects' experience of conict both
as a member of their in-group, and in opposition to an out-group. We use the MLPG
game to measure the eect of these conict manipulations upon subjects' willingness
to cooperate both within their in-group as well as between the in- and out-groups. In
particular, in our
arbitrary-groups
treatment, subjects play the MLPG game as the rst
stage of the experiment, without any prior experience of the Tullock contest.
In this
treatment, subjects play the MLPG game without any previous history of interaction
with the members of their in- or out-groups. This treatment thus constitutes a baseline
measure of cooperativeness in the absence of any interaction history.
In our
rematched-groups
treatments, subjects play the Tullock contest as the rst stage.
However, they are rematched with a new out-group before playing the MLPG game. In
these treatments, subjects have previously interacted with the other members of their
in-group, but not the out-group, prior to playing the MLPG game. This enables us to
strengthening other-regarding
in our xed-groups treatments,
identify the eect of the experience of conict in potentially
preferences toward members of the in-group.
Finally
subjects play the Tullock contest rst, and are then paired with the same out-group for
the MLPG game. As a result, they have previously interacted with both their in- and
out-groups prior to playing the MLPG game. This enables us to identify the eect of the
experience of conict in
weakening
other-regarding preferences toward the out-group.
In addition, in our xed- and rematched-groups treatments, we also vary the return on
cooperation between groups in the MLPG game, in the form of the marginal per-capita
return (M P CR) on contributions to the global account that benets the members of both
the in- and out-groups. We do this to study how our conict manipulations inuence the
extent to which subjects respond to eciency considerations.
3.1. Tullock contest
The Tullock contest (Tullock, 1980) models an unproductive conict between two parties
over an exogenous prize
P.
In our implementation of this game, we take each group to
be a party to the contest, with the prize to be contested between two groups and then
divided equally among the three members of the winning group.
group has an endowment
y,
and must choose an amount
x
In each round, each
to invest in its contest fund
to increase its chances of winning the prize. Given the investments of the two groups,
and
xh ,
the probability that group
g
is the winner is given by:
8
xg
xg
,
xg + xh
Pr (P |xg , xh ) =
and the expected payo to group
g
(1)
in a given round of the game is:
Πg = (y − xg ) +
xg
P
xg + xh
(2)
The Tullock game has a unique equilibrium (in terms of total group investments) which
can be found by taking the rst-order condition of
Πg
xg = xh = x∗
in its respective contest fund.
such that each group invests
x∗ = P/4
y = 300,
We give each group an aggregate endowment of
over a prize of the same value (i.e.
∗
x = 75
P = 300),
with respect to
xg
and setting
and the two groups compete
implying an equilibrium investment of
for each group. Since the prize is split equally among members of the winning
group, each group member receives
P/3 = 100 in the event that their group is the winner.
To conduct the Tullock game in groups while preserving the unique equilibrium, and also
to avoid wealth eects among the members of an in-group (which might inuence their
contribution decisions in a subsequent MLPG game), we determine investments in the
group contest fund using a median-voter rule. Each group member is given an endowment
of
yi = y/3 = 100,
and is obliged to invest the median of the amounts proposed by the
three members of their in-group, with no possibility to free-ride.
individual's share of the group investment is thus
In equilibrium, each
x∗ /3 = 25.
Under the median-voter rule, no individual has any incentive to deviate from proposing
their own true preferred level of investment, even where this diers from the risk-neutral
Nash investment, for example as a result of social or risk preferences.
In each round,
before the draw to determine the winning group occurs, subjects receive feedback on the
median investment proposed by the members of their own group, the resulting allocation
of their group to the contest fund, the allocation of the rival group, and their group's
resultant probability of being the winner. After this, the draw to determine the winner
takes place and subjects are informed of the result before continuing to the next round.
3.2. Multi-level public good game
In the MLPG game, each subject is given an individual endowment of
ωi = 100
in each
round. Each subject must decide how to allocate this endowment between three accounts:
a private account that benets the individual alone, a local public good that benets the
three members of the in-group only, and a global public good that benets six individuals:
the three members of the in-group as well as the three members of the out-group with
whom they have been matched. In the instructions, these three alternatives are framed
neutrally as accounts A, B and C.
9
Given the contribution decisions of all six players, the monetary payo to individual
i
in
any given round of the game is:
n
2n
αX
β X
πi = (ωi − ci − Ci ) +
cj +
Ck
n j=1
2n k=1
where
ci
denotes a contribution to the local public good that benets the
of the in-group and
2n = 6
(3)
Ci
n = 3 members
denotes a contribution to the global public good that benets the
members of both the in- and out-groups.
Allocations to the private account always yield a return of
1,
accruing to the individual
alone. In all treatments, the sum of contributions to the local public good, by all three
α = 1.5 and divided equally between
α/n = 0.5.
members of the in-group, is multiplied by a factor of
them, giving an
M P CR
from the local account of
Similarly, the sum of contributions to the global public good, by all six members of the inand out-groups, is multiplied by a factor of
β
and divided equally between them. We vary
the return to the global account between treatments. In our
treatments we set
in our
Since
high-gains
β = 2,
M P CR on the global account of β/2n = 0.33,
set β = 3, giving an M P CR of β/2n = 0.5.
giving an
treatments we
β/2n ≤ α/n < 1
low-gains-from-cooperation
while
in all treatments, an individual who cares only about her own
material payo will contribute nothing to either of the public goods, just as in a standard
(single-level) public good game. On the other hand, since
β>α>1
in all treatments,
full contribution to the global account is always the most socially ecient outcome.
3.3. Treatments
By manipulating the nature of the prior group interactions (if any) before the play of
the MLPG game, we are able to identify the eect that exposure to inter-group conict,
involving either the same or a dierent out-group, has upon preferences for both intraand inter-group cooperation, and thus examine how subjects' other-regarding preferences
are shaped by the experience of conict.
Moreover, by manipulating the gains from
cooperating with the out-group as well, we are also able to measure the response to
eciency considerations, and in particular whether this diers between treatments in
which subjects are exposed to the same, or to a dierent, out-group in the MLPG game.
In total, we have ve treatments in a
(2 × 2)+1 design.
Firstly, we interact the dimension
xed- versus rematched -groups with that of low versus high gains from cooperation.
This results in four treatments: xed groups with low / high gains from cooperation (FL
and FH, respectively) and rematched groups with low / high gains from cooperation (RL
of
10
and RH, respectively). In these four treatments, the Tullock contest is played as the rst
stage to induce a prior experience of conict before playing the MLPG game.
In addition, we also include an
arbitrary groups with low gains from cooperation treatment
(AL), in which subjects play the MLPG game rst (with low gains from cooperation)
without any prior interaction with either their in- or out-groups, and then play the
Tullock contest second (in a xed group matching). This AL treatment captures baseline
levels of cooperation with no prior experience of conict as a group.
3.4. Procedures
The experiment was conducted at the Laboratory of Experimental Economics in Prague,
Czech Republic between April 2012 and October 2013. We collected data for ten group
pairs (with six subjects each) of the MLPG game in each of the ve treatments. Subjects
were recruited using ORSEE (Greiner, 2004) from a pool of students who had registered to
participate in economic experiments. A total of 300 subjects took part in the experiment.
Of these, 57% were undergraduates and 69% were males.
The experiment was conducted entirely in English.
3
Sessions were conducted with 12-30
subjects at one time, and lasted around 75 minutes.
All subjects in any given session
were assigned to the same treatment. Two experimenters were present for each session,
with the instructions read aloud by the same experimenter in all but one of the sessions.
The experiment software was programmed using z-Tree (Fischbacher, 2007).
Upon entering the lab, subjects were randomly assigned to a computer terminal, and
instructions were both read aloud and provided in print.
4
At the beginning of the session,
subjects were informed that they would complete two tasks, but they were not told
anything about the second task until after they had completed the rst one. Subjects
were told that they would be matched into groups, and that these groups would remain
anonymous both during and after the experiment.
In all treatments except for AL, the Tullock contest was played rst, followed by the
MLPG game. The instructions for the rst game were read, and then subjects answered
a series of control questions to ensure that they understood the task. After completing
the rst game, subjects were told that they would continue to the second task, and a
3 The invitations to participate indicated clearly that the experiment would be conducted in English.
All subjects completed a series of control questions (also in English) which serves to conrm that they
were procient enough in English to understand the instructions.
4 The instructions for treatment RL are available online as Appendix B.
11
similar procedure was followed again. The instructions clearly stated whether the outgroup matching was the same or dierent from the rst task (according to the treatment),
and that the composition of the in-group would remain unchanged across both tasks.
Each subject was paid for one round of the Tullock contest and one round of the MLPG
game, chosen at random after both games had been completed.
All payments were
made in private; the average payment per subject was 250 CZK, which was equivalent to
approximately $13 USD at the time of the experiment.
4. Theory
4.1. Model
We now develop our hypotheses regarding the eects of our treatments upon the level
of contributions to the local and global public goods in the MLPG game. We motivate
these hypotheses with a very simple reduced-form model of other-regarding preferences.
In particular we allow for the possibility that an agent may, in addition to their own
material payos, attach some weight to the payos of others in the utility function.
At the same time, we allow for the possibility that the agent may distinguish between
members of the in- and out-groups in the weights that they assign to the payos of others.
Thus, let
a
denote the value that an agent attaches to an extra unit of payo to another
member of the in-group, relative to a unit of own payo, and let
attached to an extra unit of payo to a member of the out-group.
b
denote the weight
In reality,
a
and
b
are latent taste parameters that are heterogeneous across the population. However, for
concreteness we speak of the eects of our treatments upon the tastes of a representative
agent. We assume at minimum that
a, b ∈ (−1, 1), with a ≥ b.
Then, taking into account
the eect of contributions to the local and global public goods upon the payos of others,
we augment the material payo function (3) to write the utility function as:
5
n
2n
X
X
β
α
cj +
[1 + (n − 1) a + nb]
Ck
Ui = (ωi − ci − Ci ) + [1 + (n − 1) a]
n
2n
j=1
k=1
In our experiment design,
α = 1.5 and n = 3 are xed, β
(4)
is varied as a treatment variable
in our low and high gains from cooperation conditions, and we hypothesize that our
arbitrary, rematched and xed group assignment conditions will shift the other-regarding
preference parameters
a
and
b
in a manner that we describe below.
5 Clearly,
equation (4) omits payos that others derive from allocations to their own private accounts.
We do this to focus on the indirect, but partial, eects of an agent's own contributions to the local and
global public goods upon their own utility. That is, we do not claim to provide a full equilibrium model.
12
Figure 1: Model predictions
β=2
b
b
β=3
b=a
0.5
1
b=a
a
0.5
1
a
Given the linear specication of the payo and utility functions, this model generates
the highly stylized point prediction that an agent should always contribute fully to
whichever account yields the highest marginal benet respectively 1,
0.5 (1 + 2a),
and
β
6
(1 + 2a + 3b) for the private, local and global accounts. Figure 1 depicts the parameter
congurations of a and b for which it is optimal to contribute to the private (red, lower
left), local (blue, lower right) and global (green, upper right) accounts respectively. We
depict these for the case of
β = 2
in the left-hand panel, and for
β = 3
on the right.
While our model clearly oversimplies in predicting only boundary allocations, it serves
to generate some useful intuition which we illustrate with reference to some special cases.
Firstly, for the case of a purely selsh agent with
a = b = 0, the model yields the standard
prediction that own material payos are maximized by allocating the entire endowment
to the private account and contributing nothing to either the local or global public goods.
Secondly, consider the case
a>b=0
in which an agent assigns positive weight to the
payos of other members of the in-group, but is indierent to the payos of the out-group.
This corresponds to points along the horizontal axis in Figure 1. In this case, the agent
will contribute to a public good if they are suciently regarding of the other members of
a > 0.5. Then, for β = 2 it is optimal to contribute fully to the local
β = 3 the agent is indierent between the local and global accounts.
the in-group, i.e. if
account, while for
Thirdly, consider the case
a = b > 0 in which the agent assigns equal and positive weight
to the payos of all members of the in- and out-groups alike (i.e. the agent does not
perceive any meaningful distinction between the two groups). This corresponds to points
along the diagonal in Figure 1. In this case, the agent would contribute fully to the global
13
Figure 2: Hypothesized eects of treatments upon model parameters
AL
‹a (H1)
RL
Œb (H2a)
FL
‹β (H3a)
‹β (H3b)
RH
public good if
a = b > 0.4 for β = 2,
or
FH
Œb (H2b)
a = b > 0.2 for β = 3.
Otherwise, it is optimal to
retain the entire endowment in the private account. It follows that, for both values of
β
in our design, it is only optimal to contribute to the local account if the agent attaches
less weight to the payos of members of the out-group than to the in-group, i.e. if
b < a.
4.2. Hypotheses
Recall that our experiment design has two dimensions. Firstly, we manipulate the values
of the other-regarding preference parameters
rematched
and
global account
xed
a
and
b
indirectly through our
arbitrary,
group assignment conditions. Secondly, we vary the return on the
β directly in our low and high gains from cooperation conditions.
In Figure
2, we summarize the hypothesized eects of our treatments upon these three parameters.
Our rst set of hypotheses are concerned with the eects of our group matching manipulations, operating through the other-regarding preference parameters
a
and
b.
In our arbitrary groups treatment, subjects had no prior interaction with either their inor out-groups before playing the MLPG game.
By contrast, in our rematched groups
treatments, they had previously interacted with their in-group in the Tullock contest
but with a dierent out-group.
Finally, in our xed groups treatments subjects had
previously interacted with both the same in- and out-groups in the Tullock game.
We hypothesize rstly that, relative to the arbitrary groups condition, a shared experience
of conict may strengthen subjects' other-regarding preferences toward members of the
in-group, as captured by the parameter
a.
This corresponds to a rightward movement in
Figure 1, and has the eect of making contribution to the local account more attractive
under both rematched and xed groups.
We hypothesize secondly that, relative to rematched groups, a past experience of conict
with the same out-group may weaken subjects' other-regarding preferences toward members of the out-group, as captured by the parameter
14
b.
This corresponds to a downward
movement in Figure 1, and has the eect of making contribution to the global account
less attractive specically under xed groups only.
On the basis of these hypothesized eects, our model implies the following predictions:
Hypothesis 1:
Contributions to the local account will be higher under the RL and FL
treatments compared to the AL treatment.
Hypothesis 2a:
Contributions to the global account will be lower under the FL treat-
ment compared to the RL treatment.
Hypothesis 2b:
Contributions to the global account will be lower under the FH treat-
ment compared to the RH treatment.
Our next set of hypotheses are concerned with the second dimension of our experiment
design in which we vary the gains from between-group cooperation as captured by the
parameter
β.
It follows directly from equation (4) that an increase in
β
will increase
the marginal benet to contributing to the global account. However, the magnitude of
this increase depends also on the value of
b,
which we have hypothesized above to be
attenuated under our xed groups treatments. Accordingly, our model implies that:
Hypothesis 3a:
Contributions to the global account will be higher under treatment RH
compared to treatment RL.
Hypothesis 3b:
Contributions to the global account will be higher under treatment FH
compared to treatment FL. However, the magnitude of this increase will be smaller
than under rematched groups.
5. Results
In Table 1, we report summary statistics of contributions in the MLPG game.
6
For each
treatment, we compute the mean allocation to each account pooled over all ten rounds,
as well as mean earnings.
7
The gures in parentheses are treatment standard deviations,
treating group pairs as observations, i.e. we treat the mean in each group pair as a single
observation and report the standard deviation for the ten group pairs in each treatment.
6 We
8
report our analysis of the Tullock contest in Section 5.3 and Appendix Figure A1.
Figure A2 depicts the time paths of mean individual allocations to the private, local and
global accounts for each of the ve treatments. It is clear that the ranking of the treatments with respect
7 Appendix
15
Table 1: Summary statistics (by group pairs)
Treatment
Obs
Arbitrary Low
10
Rematched Low
10
Private
Local
Global
7.7
33.2
(4.1)
(15.3)
59.2
(13.8)
48.3
10.1
(23.5)
Fixed Low
10
(9.9)
46.4
15.0
(18.1)
Rematched High
Fixed High
Note:
10
10
(9.3)
41.6
(22.3)
38.6
(19.1)
36.1
5.4
58.5
(15.2)
(4.4)
(16.3)
48.0
7.5
44.5
(15.0)
(4.0)
(14.6)
Earnings
137.0
(14.4)
146.6
(22.4)
146.1
(18.0)
219.6
(31.9)
192.7
(29.1)
standard deviations in parentheses.
Table 2: Wilcoxon rank-sum p-values (two-sided, by group pairs)
Private
Local
0.940
Global
H1: RL vs. AL
0.406
0.326
H1: FL vs. AL
0.049**
0.041**
0.496
H2a: FL vs. RL
0.821
0.290
0.821
H2b: FH vs. RH
0.112
0.290
0.059*
H3a: RH vs. RL
0.227
0.290
0.041**
H3b: FH vs. FL
0.762
0.028**
0.406
Note:
*
p < 0.10;
**
p < 0.05.
Table 1 indicates that contributions to the global account are highest, while allocations to
the private account are lowest, under treatment RH. On the other hand, contributions to
the global account are lowest, while allocations to the private account are highest, under
treatment AL. Contributions to the local account are low across all ve treatments, with
the highest level (at 15%) observed under treatment FL. The ranking of the treatments
with respect to earnings is identical to the ranking with respect to global contributions,
even though the treatments dier with respect to the eciency of the global public good,
and there is also the opportunity to contribute to the less ecient local account.
In Table 2, we report two-sided
p-values
for Wilcoxon rank-sum tests of the equality of
contributions, to each of the three accounts, in each pairwise comparison between treatments highlighted by our hypotheses. Again, this analysis treats the mean contribution
of each group pair by all six subjects and in all ten rounds as a single observation.
to the level of contributions to each of the accounts is fairly stable over the ten rounds of the MLPG game,
and there are no obvious dierences in either the nature or slope of the time trends across treatments.
For these reasons, we aggregate the data from all ten rounds throughout our analysis.
8 We acknowledge that in treatments with rematched groups, the group pairs are not strictly speaking
independent observations. This is because subjects have previously interacted with members of a dierent
out-group in the rst-stage Tullock contest, and this could potentially result in inter-dependencies in
behavior across two group pairs.
16
Table 3: Tobit regressions of mean individual contributions
Private
H1: RL
-11.676
(8.972)
Global
2.940
8.575
(4.278)
(8.739)
H1: FL
-13.003*
RH
-23.805***
(6.878)
(3.100)
(7.298)
FH
-11.362*
-1.844
11.550*
(6.750)
(2.630)
(6.823)
(7.376)
Observations
Notes:
Local
9.662**
5.622
(3.899)
(7.850)
-4.825
26.238***
300
300
300
H2a: FL vs. RL
0.889
0.209
0.753
H2b: FH vs. RH
0.076*
0.388
0.039**
H3a: RH vs. RL
0.186
0.113
0.048**
H3b: FH vs. FL
0.826
0.007***
0.437
base category AL; robust standard errors clustered by group pairs.
*
p < 0.10;
**
p < 0.05;
***
p < 0.01.
Finally, in Table 3 we report an individual-level regression analysis of the eects of our
treatments upon individual contributions to each of the three accounts. For each of the
300 subjects, we compute the mean amount allocated by that subject to each account
over the ten rounds of the MLPG game. We regress these mean contributions on a set
of dummies for each of the treatments, in a two-limit Tobit model with treatment AL
as the omitted category. Each subject contributes one observation to each of the three
regressions, and we report robust standard errors clustered at the level of group pairs.
Table 3 also reports two-sided
p-values for tests of the equality of the coecients for each
of the pairwise comparisons between treatments highlighted by our hypotheses.
On the basis of these analyses, we report our main results in the following two subsections.
5.1. Eects of group matching
Our rst two hypotheses are concerned with the eects of our
matched
arbitrary, xed
and
re-
groups manipulations, and correspond to the horizontal dimension in Figure
2. We summarize the eects of this dimension of our experiment design graphically in
Figure 3, separately for each of the three accounts, and for treatments with low and high
gains from cooperation. The condence bars in this gure represent
±1
standard error
of the mean, treating group pairs as observations.
Hypothesis 1 states that subjects' other-regarding preferences toward the members of
their in-group may be strengthened when they have had the shared experience of playing
the Tullock contest together.
As a result, we expect contributions to the local public
good to be higher under treatments RL and FL compared to treatment AL.
17
Figure 3: Eects of group matching
A. Effect of Group Matching with Low Gains to Cooperation
Private
Local
Global
70
60
50
40
30
20
10
0
AL
RL
FL
AL
RL
FL
AL
RL
FL
B. Effect of Group Matching with High Gains to Cooperation
Private
Local
Global
70
60
50
40
30
20
10
0
RH
FH
RH
FH
RH
FH
Note: confidence bars are +/− 1 SEM, treating group pairs as observations.
18
Both the graphical presentation in Figure 3 as well as the summary statistics in Table 1
conrm that our results are directionally consistent with these predictions: contributions
to the local account are highest under FL (15.0%) followed by RL (10.1%) and AL (7.7%).
This comes at the expense of allocations to the private account, which are higher under AL
(59.2%) than RL (48.3%) and FL (46.4%). Tables 2 and 3 indicate that the dierence in
contributions to the local account between FL and AL is statistically signicant at the 5%
level both in a nonparametric test at the level of group pairs, as well as in the individuallevel regression.
9
On the other hand, none of the dierences between treatments RL and
AL are signicant in any of the analyses.
In interpreting these results, we acknowledge that the comparison between treatments
FL and AL is not as clean as the one between RL and AL. Since subjects in FL also
previously competed with the same out-group in the Tullock contest, the dierences
that we observe may also reect an eect of negative sentiment toward the out-group as
implied by Hypothesis 2. We summarize our discussion of Hypothesis 1 as follows:
Result 1:
Contributions to the local public good are signicantly higher under treatment
FL compared to AL. Subjects are more cooperative toward the members of their
in-group when they have previously jointly competed against the same out-group,
compared to when they have not previously interacted with the members of either
group. This result likely reects a combination of in- and out-group eects.
Hypothesis 2 states that subjects' other-regarding preferences toward the members of
their out-group may be weakened when they have previously competed against the same
out-group in the Tullock contest. As a result, we expect contributions to the global public
good to be lower under treatment FL compared to RL, as well as in FH compared to RH.
Once again, both Figure 3 and Table 1 conrm that our results are directionally consistent
with these predictions:
contributions to the global account are lower both under FL
(38.6%) compared to RL (41.6%), as well as under FH (44.5%) compared to RH (58.5%).
Tables 2 and 3 indicate that when gains from cooperation are high, the dierence in
contributions to the global account between FH and RH is signicant at the 5% level
in the individual-level regression, and at the 10% level in the rank-sum test at the level
of group pairs.
10
On the other hand, when gains from cooperation are low, none of the
dierences between treatments FL and RL are signicant in any of the analyses.
9 The
osetting dierence in allocations to the private account is also signicant at the 5% level in
the rank-sum test in Table 2, although it is only marginally signicant at the 10% level in the regression
model of Table 3.
10 In addition, allocations to the private account are higher in FH compared to RH, and this dierence
is marginally signicant at the 10% level in the individual-level regression model.
19
As a result of their reluctance to cooperate with the members of an out-group with
whom they have previously competed in the Tullock contest, subjects in our xed groups
treatments attain lower earnings and hence a lower level of eciency. When the gains
from cooperation are low, these costs are small and not statistically signicant: average
earnings drop fractionally from 146.6 under RL (out of a maximum of 200 in the low
gains treatments, implying an eciency of 73.3%) to 146.1 under FL. This dierence is
clearly not signicant, with
p = 0.880
in a Wilcoxon rank-sum test.
However, when the gains from cooperation are large, the costs are more substantial:
average earnings drop from 219.6 under RH (out of a maximum of 300 in high gains
treatments, implying an eciency of 73.2%) to 192.7 (64.2%). It turns out that the eciency of the FH treatment is the lowest out of our ve treatments. In an OLS regression,
analogous to the Tobit models in Table 3, in which we regress each subject's mean earnings over the ten rounds of the MLPG game on dummies for each of the treatments, with
standard errors clustered at the level of group pairs, we nd the dierence in earnings
between FH and RH to be signicant with
p = 0.047.11
We also nd this dierence to be
marginally signicant in a Wilcoxon rank-sum test, with
Result 2:
p = 0.070.
When the gains to cooperation between groups are large, contributions to the
global public good are signicantly lower under treatment FH compared to RH.
Subjects are less cooperative toward the members of their out-group when they
have previously competed against that group, compared to when their previous
interaction was with a dierent out-group.
As a result of this out-group bias,
subjects attain signicantly lower levels of earnings and eciency.
One potential concern with our interpretation of these group matching eects is that experience of the Tullock contest could be informative to subjects regarding the preferences
of both in- and out-group members, and this could inuence behavior in the MLPG game
independently of the hypothesized eects upon their preferences. In particular, subjects
in the xed matching treatments might form more accurate beliefs as a result of having
previously interacted with both their in- and out-groups. In this event, we might expect
dierences between treatments to decrease over the ten rounds of the MLPG game, as
subjects in the rematched groups treatments learn the preferences of their new out-group.
However, as can be seen from Appendix Figure A2, this is not what we observe: the differences between treatments in contributions to the local and global public goods remain
stable over time, with all treatments displaying very similar time trends.
11 Full
results are available upon request.
20
5.2. Eects of gains from cooperation
Our nal hypothesis is concerned with the eects of our low versus high gains from
cooperation manipulation, and corresponds to the vertical dimension in Figure 2.
We
summarize these eects graphically in Figure 4, separately for each of the three accounts,
and for treatments with rematched and xed groups. Once again, the condence bars in
this gure represent
±1 standard error of the mean, treating group pairs as observations.
Hypothesis 3 states that an increase in the return on contributions to the global public good, representing the magnitude of potential gains from cooperation with the outgroup, will increase the attractiveness of contributing to the global account. However,
the response to this increase depends also on the strength of subjects' other-regarding
preferences toward their out-group, which are also hypothesized to be weakened when
the two groups have previously competed in the Tullock contest. Accordingly, we expect
contributions to the global public good to be higher under treatment RH compared to RL.
We also expect global contributions to be higher under FH compared to FL, however we
expect this latter eect to be smaller in comparison to the rematched groups treatments.
Both Figure 4 and Table 1 conrm that our results are directionally consistent with these
predictions: contributions to the global account are higher both under RH (58.5%) compared to RL (41.6%), as well as under FH (44.5%) compared to FL (38.6%). Moreover,
the dierence under rematched groups (16.9%) is almost three times larger than under
xed groups (5.9%). As a result, Tables 2 and 3 indicate that under rematched groups,
the dierence in contributions to the global account between RH and RL is statistically
signicant at the 5% level both in a nonparametric test at the level of group pairs, as
well as in the individual-level regression.
On the other hand, under xed groups the dierence in global contributions between FH
and FL is not signicant in either of the analyses. Thus when there is a prior history of
conict between the two groups, subjects appear to be largely unmoved by an increase
in the return to cooperating with the out-group.
Result 3:
Contributions to the global public good are signicantly higher under treat-
ment RH compared to RL. Subjects respond to an increase in the return to cooperating with the out-group when the two groups have not previously interacted.
However, when the two groups have previously competed against one another, there
is no signicant response.
Finally, we note one result not predicted by our model: contributions to the local public
good are higher under FL compared to FH, and this dierence is clearly signicant at
the 5% level in the Wilcoxon test, and at the 1% level in the regression analysis.
21
Figure 4: Eects of gains from cooperation
A. Effect of Gains from Cooperation in Rematched Groups
Private
Local
Global
70
60
50
40
30
20
10
0
RL
RH
RL
RH
RL
RH
B. Effect of Gains from Cooperation in Fixed Groups
Private
Local
Global
70
60
50
40
30
20
10
0
FL
FH
FL
FH
FL
FH
Note: confidence bars are +/− 1 SEM, treating group pairs as observations.
22
5.3. Tullock contest
Since our group matching treatments were intended to manipulate subjects' experience
of conict, we now verify that the Tullock games were indeed keenly contested. Appendix
Figure A1 plots the mean individual proposed investment in the group contest fund for
each round of the Tullock game.
12
In our design, the actual investment that was binding
on each subject is the median of the amounts proposed by the three members of their
in-group; Figure A1 also plots the mean of this group median.
Recall that each subject has an endowment of 100 in each round, and that the risk-neutral
Nash investment at the individual level is 25. Figure A1 indicates that the mean proposed
investments are substantially higher, and typically around double the Nash level. Further,
the group medians are on average very close to the individual means, indicating that the
high mean proposals are not driven by outlying group members. There is no discernible
time trend in the level of contest investments over time, and in particular there is no
evidence of convergence toward the Nash investment.
To give an indication of the extent of heterogeneity in the intensity of conict in dierent
group pairs, we compute the mean of the six individual proposals within each group pair,
and depict the inter-quartile range of these means with respect to the forty group pairs
as the shaded region in Figure A1. This conrms that while there are indeed dierences
in the intensity of conict between group pairs, even the comparatively less competitive
group pairs nonetheless invest at substantially higher than the Nash level.
5.4. Discussion
Our results demonstrate that competition does indeed make group membership more
salient than arbitrary group identity in a laboratory setting. Contributions to the local
account are signicantly higher in the FL treatment compared to AL (Result 1).
In-
terestingly, while the eect is directionally the same when comparing RL to AL, the
dierence is not statistically signicant. Since decisions in the FL treatment are aected
by preferences toward both in- and out-group members, the dierence in allocations to
the local account in FL compared to AL may be driven in part by a decrease in generosity
toward an out-group with whom one has previously been in conict (i.e. a decrease in
in equation 4, as well as an increase in
b
a).
Another intriguing possibility is that dierent preferences for cooperation with the ingroup may be activated when continuing to interact with the same out-group as in the
12 This
gure omits the data of the AL treatment, in which subjects played the Tullock contest as the
second stage, following the MLPG game.
23
contest. In terms of our model, the increase in
a
following conict may be dependent on
whether the in-group interacts with the same, or a dierent, out-group. Either of these
mechanisms would suggest that inter-group competition is a useful tool for inducing
salient group identity in economic experiments: the Tullock contest produced behavior
that diered signicantly from arbitrary group identity.
This opens the possibility of
using this, or other, forms of inter-group competition to examine questions relating to
group identity in economic experiments.
While we observe greater cooperation within groups following conict, there are substantial negative eects on eciency as a result of reduced cooperation between groups when the MLPG game is played with the same out-group as the Tullock game. In our
high returns from cooperation condition, contributions to the global account are lower
in treatment FH where the groups previously competed against one another compared to RH (Result 2). Moreover, in our xed-matching condition we nd no signicant
increase in allocations to the global account in treatment FH compared to FL. On the
other hand, allocations to the global public good increase signicantly in the corresponding rematched-groups treatments (Result 3), which is consistent with previous research
on the MLPG game using milder forms of group identity (Blackwell and McKee, 2003;
Fellner and Lünser, 2008). Taken together, these results indicate that the ndings of the
previous studies may not be robust to the form of the group identity manipulation.
The muted response to higher returns from cooperation in our xed matching treatments
has important implications for naturally-occurring conict. In our experiment, conict
in the Tullock contest is socially inecient. However in other settings, competition may
help to achieve socially ecient outcomes.
For example, in a laboratory experiment
Nalbantian and Schotter (1997) nd that among several schemes commonly used by
employers to incentivize workers, creating competition among teams is the most ecient
(see also Guillen et al., in press).
However, if conict also decreases the potential for
teams to cooperate with each other in future, the overall eect of incentivizing intergroup competition may well be negative. As a result of their prior history of competition,
changes in preferences toward (or beliefs about) members of a formerly-competing team
may impede future cooperation, resulting in lower prots for the rm in the long run.
Behavior in public good games has been explained by several mechanisms, including altruism (Andreoni, 1990), aversion to inequality (Fehr and Schmidt, 1999; Bolton and
Ockenfels, 2000), preferences for social eciency (Charness and Rabin, 2002) and reciprocity (Rabin, 1993; Dufwenberg and Kirchsteiger, 2004; Falk and Fischbacher, 2006).
Our model is agnostic on which of these mechanisms drives the eects that we observe, in
the sense that one or more of these forces may be responsible for shifting the reduced-form
utility weights,
a
and
b,
that an individual attaches to the payos of others. Moreover,
24
while our results are consistent with previous research showing that group identity aects
both in- and out-group preferences, the form of conict that we introduce in our design,
namely competition over a xed resource, may operate through distinct behavioral channels to other forms of salient group identity.
While more research is needed to explore this issue, our ndings are also broadly consistent
with eld experiments such as Gumen (2012) and Buchan et al. (2009) who nd ingroup biases in similar designs using naturally-occurring group identities, as well as with
previous work suggesting a link between parochialism and violent conict (Choi and
Bowles, 2007; Voors et al., 2012; Bauer et al., 2014). The fact that we nd a similar eect
in a neutrally-framed laboratory setting suggests that the mere act of competition over a
xed-resource increases in-group bias, even in the absence of underlying ethnic divisions,
cultural stereotyping, or exposure to violence.
6. Conclusion
In this paper, we present the results of a laboratory experiment in which we manipulate
the nature of subjects' prior exposure to conict, to study its eects upon subsequent
cooperation both within and between groups. Our design introduces a novel method to
induce a stronger form of group identity in the lab, which enables us to disentangle the
role of conict in strengthening in-group identity from its eect in changing preferences
towards an out-group. We also examine the response to changes in the returns to intergroup cooperation when there has been a past history of conict between the groups.
We nd that group identity is indeed strengthened by exposure to the Tullock contest,
and that subjects demonstrate stronger in-group preferences when there has been a shared
history of conict between the in- and out-groups. We also nd that prior exposure to
conict involving a specic out-group matters independently of the common in-group
experience of conict.
Moreover, we nd no response to an increase in the returns to
between-group cooperation when there has been a previous history of conict involving
the same out-group.
This neatly demonstrates how inter-group conict even in the
setting of a laboratory experiment can lead to less socially ecient outcomes.
Our results are consistent with a simple model in which an individual's other-regarding
preferences are sensitive to group identity, such that increases in the material payos
of in-group members may be weighted more heavily than corresponding increases in the
payos of the out-group. We nd that a shared experience of conict with one's in-group
increases the weight attached to in-group payos, while a history of conict involving a
specic out-group decreases the utility of out-group payos. This implies that conict
25
increases parochialism both by increasing preferences for in-group cooperation, and also
by decreasing preferences for out-group cooperation.
Our ndings are also consistent with those of several eld experiments using naturallyoccurring group identity and conict. The fact that we observe similar eects suggests
that competition itself plays a role in forming group identity, independent of more deeplyseated sources of group aliation and conict.
26
References
Abbink, K., 2012. Laboratory experiments on conict, in: Garnkel, M.R., Skaperdas, S. (Eds.),
The Oxford Handbook of the Economics of Peace and Conict. Oxford University Press, New
York.
Abbink, K., Brandts, J., Herrman, B., Orzen, H., 2010.
Intergroup conict and intra-group
punishment in an experimental contest game. American Economic Review 100, 420447.
Andreoni, J., 1990.
Impure altruism and donations to public goods: A theory of warm-glow
giving. Economic Journal 100, 464477.
Bauer, M., Cassar, A., Chytilova, J., Henrich, J., 2014. War's enduring eects on the development of egalitarian motivations and in-group biases. Psychological Science 25, 4757.
Blackwell, C., McKee, M., 2003. Only for my own neighborhood? Preferences and voluntary
provision of local and global public goods. Journal of Economic Behavior and Organization
52, 115131.
Bolton, G.E., Ockenfels, A., 2000.
ERC: A theory of equity, reciprocity, and competition.
American Economic Review 90, 166193.
Bowles, S., Gintis, H., 2011.
A Cooperative Species: Human Reciprocity and Its Evolution.
Princeton University Press, Princeton.
Brunwasser, M., 2011. Bosniaks and Croats, divided in class and at play. International Herald
Tribune July 2.
Buchan, N., Grimalda, G., Wilson, R., Brewer, M., Fatas, E., Foddy, M., 2009. Globalization
and human cooperation. Proceedings of the National Academy of Sciences 106, 41384142.
Chakravarty, S., Fonseca, M., 2013.
Discrimination via exclusion:
An experiment on group
identity and club goods. University of Exeter, Economics Department Discussion Paper 13/02.
Charness, G., Rabin, M., 2002. Understanding social preferences with simple tests. Quarterly
Journal of Economics 117, 817869.
Charness, G., Rigotti, L., Rustichini, A., 2007.
Individual behavior and group membership.
American Economic Review 97, 13401352.
Chen, Y., Li, S.X., 2009. Group identity and social preferences. American Economic Review 99,
431457.
Choi, J., Bowles, S., 2007. The coevolution of parochial altruism and war. Science 318, 636640.
Dufwenberg, M., Kirchsteiger, G., 2004. A theory of sequential reciprocity. Games and Economic
Behavior 47, 268298.
Eckel, C., Grossman, P., 2005. Managing diversity by creating team identity. Journal of Economic Behavior and Organization 58, 371392.
Falk, A., Fischbacher, U., 2006.
A theory of reciprocity.
Games and Economic Behavior 54,
293315.
Fehr, E., Schmidt, K.M., 1999. A theory of fairness, competition, and cooperation. Quarterly
Journal of Economics 114, 817868.
27
Fellner, G., Lünser, G., 2008. Cooperation in local and global groups. WU Vienna University
of Economics and Business, Department of Economics Working Paper 122.
Fischbacher, U., 2007. z-Tree: Zurich toolbox for ready-made economic experiments. Experimental Economics 10, 171178.
Gneezy, A., Fessler, D., 2012. Conict, sticks and carrots: War increases prosocial punishments
and rewards. Proceedings of the Royal Society B: Biological Sciences 279, 219223.
Greiner, B., 2004.
An online recruitment system for economic experiments, in: Kremer, K.,
Macho, V. (Eds.), Forschung und wissenschaftliches Rechnen 2003. Gesellschaft für Wissenschaftliche Datenverarbeitung, Göttingen, pp. 7993.
Guillen, P., Merrett, D., Slonim, R., in press. A new solution for the moral hazard problem in
team production. Management Science .
Gumen, A., 2012. Ethnic conict and the provision of public goods: A framed eld experiment.
Mimeo .
Halevy, N., Weisel, O., Bornstein, G., 2012.
In-group love and out-group hate in repeated
interactions between groups. Journal of Behavioural Decision Making 25, 188195.
Hirshleifer, J., 1995. Theorizing about conict, in: Hartley, K., Sandler, T. (Eds.), Handbook
of Defense Economics. North Holland, pp. 166189.
Ke, C., Konrad, K.A., Morath, F., 2013.
Brothers in arms An experiment on the alliance
puzzle. Games and Economic Behavior 77, 6176.
Nalbantian, H.R., Schotter, A., 1997.
Productivity under group incentives: An experimental
study. American Economic Review 87, 314341.
Puurtinen, M., Mappes, T., 2009. Between-group competition and human cooperation. Proceedings of the Royal Society B: Biological Sciences 276, 355360.
Rabin, M., 1993. Incorporating fairness into game theory and economics. American Economic
Review 83, 12811302.
Sheremeta, R.M., 2013. Overbidding and heterogeneous behavior in contest experiments. Journal
of Economic Surveys 27, 491514.
Tajfel, H., 1971.
Social categorization and intergroup behavior.
European Journal of Social
Psychology 1, 148179.
Tajfel, H., Turner, J.C., 1979.
An integrative theory of intergroup conict, in: Austin, W.,
Worchel, S. (Eds.), The social psychology of intergroup relations. Brooks/Cole, Monterey,
CA, pp. 3347.
Tullock, G., 1980.
Ecient rent-seeking, in: Buchanan, J., Tollison, R., Tullock, G. (Eds.),
Towards a Theory of the Rent-Seeking Society. Texas A&M University Press, pp. 97112.
Voors, M., Nillesen, E., Verwimp, P., Bulte, E., Lensink, R., Soest, D.V., 2012. Violent conict
and behavior: A eld experiment in Burundi. American Economic Review 102, 94164.
28
Appendix A. Additional analyses
Figure A1: Time path of Tullock contest investments (excluding treatment AL)
70
60
50
40
30
20
1
2
3
4
5
6
7
8
9
10
Round
Individual Proposals
Group Medians
Nash Prediction
Note: shaded region is the interquartile range of the mean individual proposal, by group pairs.
Figure A2: Time paths of MLPG contributions, by account and treatment
Mean allocation by treatment
Private
Local
Global
100
100
100
80
80
80
60
60
60
40
40
40
20
20
20
0
0
0
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
Round
AL
RL
FL
29
RH
FH
Appendix B. Instructions for the RL treatment (for online publication only)
1. General instructions
Welcome to this session. From now on, please do not talk to the other participants, or communicate with them in any other way. Mobile phones must also be switched o. If you have a
question, please raise your hand and one of us will come to you and assist you in private. These
rules are important. If you break any of these rules, we will cancel the session and dismiss all of
you without any payment.
In this experiment you will make a number of decisions. These instructions explain the decisions
that you will make and their consequences. Depending on your decisions you will earn money
which will be paid to you in cash at the end of the experiment.
Throughout the experiment we will record all earnings in tokens . At the end of the session, we
will randomly select two decision rounds to count toward your earnings. The tokens that you
earn in these two rounds will be converted into Czech crowns at the following exchange rate:
1 token = 1 CZK
2. First task
This task will consist of 10 rounds. At the end of the session, we will randomly draw a number
between 1 and 10 from a bag to select one of these rounds, and you will be paid your earnings
from this randomly chosen round.
At the beginning of this task you will be matched with two other randomly selected people in
the room, to form a group of three. Your group will play against one of the other groups who
will be your opponents. The other members of your group, as well as your opponents, remain
the same through all ten rounds of this task. You will not learn who your group members or
opponents are, either during or after today's session. Likewise, neither your group members nor
opponents will learn of your identity.
In each round your group and your opponents will compete for a prize as we will now explain.
At the beginning of each round, you will be given an endowment of 100 tokens. Each group must
decide how many tokens to allocate to its contest fund . This decision is made in the following
way. Each group member will be asked to propose a number of tokens to allocate to the contest
fund. The computer will then determine the median amount proposed by the three members.
(The median is the middle number of an increasing series of numbers: the median of 1, 2 and 3
is 2; the median of 1, 98 and 100 is 98.) This amount will be automatically deducted from each
member's endowment and allocated to the group's contest fund. Any tokens not allocated to
the contest fund will be yours to keep. Since each group member must allocate the same amount
to the contest fund, each member will end up with the same balance of tokens. Likewise, your
opponents will decide how many tokens to allocate to their contest fund in exactly the same
way.
After each group has chosen its allocation to the contest fund, the computer will conduct a
random draw to determine whether your group or your opponents win the prize. The prize is
30
worth 300 tokens, which is divided equally among the members of the group that wins it (100 for
each member). Your group's chances of winning depend on how many tokens are in its contest
fund. This works as follows: imagine that each token allocated to the contest fund by your group
and by the other group are placed in a bag, and then one token is randomly drawn from this
bag. If the token that is drawn belongs to your group, then your group wins the prize. If the
token belongs to the other group, then the other group wins the prize. Each group's chances of
winning depend on the number of tokens that it has allocated relative to the number of tokens
allocated by the other group.
For instance, if your group and your opponents each allocate the same amount to the contest
fund, each group has the same number of tokens in the bag and an equal chance of winning (a
1/2 chance). If your group allocates twice as many tokens to its contest fund as your opponents,
your group has twice the chances of winning (your group has a 2/3 chance of winning and the
other group has a 1/3 chance). Thus, your group's chances of winning increase with the amount
that it allocates to the contest fund, and decrease with the amount allocated by your opponents.
If your group allocates nothing and the other group allocates 1 or more tokens, than the other
group automatically wins. If neither group allocates anything to the contest fund, then both
groups have a 1/2 chance of winning.
After the computer has determined the winner, you will be informed which group won the prize
and shown your earnings for that round. Your earnings are equal to your initial endowment of
100 tokens, minus the number of tokens you allocated to the contest fund, plus 100 tokens if
your group won the prize. Since each group member must allocate the same number of tokens
to the contest fund in each round, each group member's earnings will also be the same in each
round.
3. Second task
This task will again consist of 10 rounds. At the end of the session, we will randomly draw a
number between 1 and 10 from a bag to select one of these rounds, and you will be paid your
earnings from this randomly chosen round.
For this task, you will again be a member of the same group of three people with whom you
were matched in the previous task. This group will now be paired with a second group of three,
who were also matched with one another in the previous task, and who we will refer to as the
other group . As before, you will never know the identities of any of these people, and they
will never know your identity.
The other group will not be the same as the group that was your opponent in the
rst task.
At the beginning of each round, you will be given an endowment of 100 tokens. You will be asked
to decide how many of these tokens you will allocate to three accounts: Account A, Account B
and Account C. Your total earnings will depend on the amount that you and others allocate to
each of the accounts as explained below:
31
ACCOUNT B:
Each member of your group earns 1/2 for each token allocated by any member of your group.
YOUR
GROUP
A
ACCOUNT A:
You earn 1 for each
token you allocate.
YOU
A
ACCOUNT C:
Each member of both groups earns 1/3 for each token allocated by any member of both groups.
OTHER
GROUP
A
A
A
B
Your earnings from Account A
Each token that you allocate to Account A will earn one token for you alone. Therefore if you
allocate X tokens to Account A, you will earn exactly X tokens from Account A.
No-one other than you earns anything from the tokens you allocate to Account A. Likewise, you
will not earn anything from any tokens allocated to Account A by any other person.
Your earnings from Account B
Tokens allocated to Account B only aect the earnings of the three members of your own group.
For every token allocated to Account B, by any member of your group, each member of your
group will earn 1/2 tokens, regardless of whether he or she allocated any tokens to Account B.
Your earnings from Account B = 1/2 x (sum of three allocations to Account B)
The earnings from Account B are calculated in the same way for all three members of your
group, so all members of your group each receive the same earnings from Account B as you do.
Therefore, all members of your group each benet equally from every token that any member
allocates to Account B.
For each token that you allocate to Account A, you earn one token. Suppose that you allocate
this to Account B instead.
Then the total amount allocated to Account B increases by one
token, and your earnings from Account B increase by 1/2 tokens.
At the same time, the earnings of the other members of your group also increase by 1/2 tokens
each, so the total earnings from Account B for all three group members would increase by 3/2
tokens in total.
32
Your allocation to Account B therefore increases the earnings of the other two members of your
group, and similarly their allocations to Account B also increase your earnings. For each token
that another member of your group allocates to Account B, you also earn 1/2 tokens.
So, if all 3 members of your group allocate 1 token to Account B, then Account B contains 3
tokens in total, and each group member will receive 3/2 tokens. In total, your group would earn
3 x 3/2 = 9/2 tokens from Account B.
Tokens allocated to Account B only aect the earnings of the members of your own group. They
do not aect the earnings of the other group of three with whom your group is matched.
The members of the other group can allocate tokens to their own Account B, and this will not
aect the earnings of you or the other members of your own group.
Your earnings from Account C
Tokens allocated to Account C aect the earnings of both the three members of your own group,
as well as the three members of the other group. For every token allocated to Account C, by
any member of either group, each member of both groups will earn 1/3 tokens, regardless of
whether he or she allocated any tokens to Account C.
Your earnings from Account C = 1/3 x (sum of six allocations to Account C)
The earnings from Account C are calculated in the same way for all six members of both groups,
so all members of both groups each receive the same earnings from Account C as you do.
Therefore, all members of both groups each benet equally from every token that any member
of either group allocates to Account C.
For each token that you allocate to Account A, you earn one token. Suppose that you allocate
this to Account C instead.
Then the total amount allocated to Account C increases by one
token, and your earnings from Account C increase by 1/3 tokens.
At the same time, the earnings of the other members of both groups also increase by 1/3 tokens
each, so the total earnings from Account C for all six members of both groups would increase
by 2 tokens in total.
Your allocation to Account C therefore increases the earnings of the other ve members of both
groups, and similarly their allocations to Account C also increase your earnings. For each token
that another member of either group allocates to Account C, you also earn 1/3 tokens.
So, if all 6 members of both groups allocate 1 token to Account C, then Account C contains 6
tokens in total, and each member of both groups will receive 2 tokens. In total, both groups
would earn 6x2=12 tokens from Account C.
Your total earnings from this task
Your earnings = Your own allocation to Account A
+ 1/2 x (sum of three allocations to Account B)
+ 1/3 x (sum of six allocations to Account C)
33